Fake_or_Real_Competition_Dataset


2023 Fake or Real: AI-generated Image Discrimination Competition dataset is now available on Hugging Face!
---

Hello🖐️

We are excited to announce the release of the dataset for the 2023 Fake or Real: AI-generated Image Discrimination Competition. The competition was held on AI CONNECT(https://aiconnect.kr/) from June 26th to July 6th, 2023, with 768 participants.

If you're interested in evaluating the performance of your model on the test dataset, we encourage you to visit the [competition page](https://aiconnect.kr/competition/detail/227/task/295/taskInfo) on AI CONNECT and submit your results. Please note that it supports only Korean yet. Of course we data scientists can always use Chrome translate, and/or even better translation models🥳. Plus, multilingual service will be provided in the (hopefully near) future, so please stay tuned!

# Background
As the advancement of generative AI technology has enabled the easy creation of indistinguishable fake information from genuine content, concerns regarding its misuse have surfaced. Image generation AI, in particular, has raised significant alarm due to its potential risks such as identity theft, revenge porn, and political manipulation. In response, it has become imperative to develop technologies that can effectively discern between real and AI-generated fake images.

The training dataset consists of diffusiondb (https://huggingface.co/datasets/poloclub/diffusiondb) and Flickr images, with the inclusion of some low-quality fake images. For the test dataset, we took measures to construct it in a manner that closely resembles real-world scenarios involving image misuse. We utilized multiple generative AI models, fine-tuned on diverse photorealistic datasets, and applied negative prompt keywords like 'cartoon' and 'too many fingers' to generate realistic images.

We hope this dataset encourages the development of robust solutions and stimulates discussions on tackling the challenges associated with AI-generated fake images.

Best Regards,
AI CONNECT


2023年「真伪辨：AI生成图像甄别」竞赛数据集现已于Hugging Face平台发布！
---

您好👋

我们欣喜地宣布，2023年「真伪辨：AI生成图像甄别」竞赛的数据集正式发布。本次竞赛于2023年6月26日至7月6日在AI CONNECT平台（https://aiconnect.kr/）举办，共有768名参赛者参与。

若您希望在测试集上评估模型的性能，欢迎访问AI CONNECT平台上的[竞赛页面](https://aiconnect.kr/competition/detail/227/task/295/taskInfo)并提交结果。请注意，目前该页面仅支持韩语界面，不过数据科学家们可以借助Chrome浏览器的翻译功能，或是更优质的翻译模型来解决这一问题🥳。此外，我们有望在不久的将来提供多语言服务，敬请期待！

# 研究背景
随着生成式AI(Generative AI)技术的发展，人们可以轻松生成与真实内容难以区分的虚假信息，由此引发了人们对其被滥用的担忧。其中，图像生成AI因可能带来身份盗用、报复性色情内容传播、政治操纵等风险，引发了广泛警惕。对此，研发能够有效甄别真实图像与AI生成虚假图像的技术已成为当务之急。

本次训练数据集由diffusiondb（https://huggingface.co/datasets/poloclub/diffusiondb）与Flickr图像构成，同时纳入了部分低质量的虚假图像。对于测试数据集，我们采取了贴近真实图像滥用场景的构建方式：使用多款在多样化写实数据集上微调后的生成式AI模型，并添加“卡通风格”“手指数量过多”等负向提示词来生成贴近真实场景的虚假图像。

我们希望本数据集能够推动鲁棒性甄别方案的研发，并激发围绕解决AI生成虚假图像相关挑战的讨论。

此致，
AI CONNECT